Models¶
Based on my previous analyses and feature selection, I can frame several distinct model questions that leverage different modeling approaches:
Model 1: Predicting Stock Price Movements with Supervised Learning¶
I plan to utilize historical stock data and financial indicators to predict future closing prices of stocks. Our independent variables will include 'Volume', 'MACD', 'Reported EPS', 'Surprise(%)', 'Revenue', and historical stock prices. I will experiment with including and excluding 'RSI' to assess its impact on predictive performance. Our dependent variable will be the 'Close' price of the stock on trading for the next quarter. I will employ a regression-based supervised learning model, such as a Random Forest Regressor, to capture the non-linear relationships observed in our exploratory data analysis. Our EDA supports this approach, as it revealed strong correlations between these features and the closing price. I will split our data into training and testing sets to evaluate the model's performance and ensure it generalizes well to unseen data.
Model 2: Optimizing Stock Trading Decisions with Contextual Bandits¶
I plan to deploy a Contextual Bandits model to optimize trading decisions in the stock market. State, Actions and Reward is defined as following:
State (Context): The state consists of stock features of 'Close' price, 'Volume', 'MACD', 'Reported EPS', 'Surprise(%)', and 'Revenue'.
Actions: The actions is "buy", "sell", or "hold".
Reward: The reward function is the immediate profit or loss from an action. For example, if the agent decides to buy, and the stock price increases in the next time step, the reward is positive. Conversely, if the price decreases, the reward is negative.
Our EDA supports this approach, as it revealed strong correlations between these features and the closing price. I will split our data into training and testing sets. I will train the model using training data and evaluate its performance using a testing data.
Another Better Model maybe considered: Optimizing Stock Trading Decisions with Deep Reinforcement Learning (DRL)¶
Deep Reinforcement Learning (DRL) model maybe a better model to consider. The state space will consist of features such as 'Close' price, 'Volume', 'MACD', 'Reported EPS', 'Surprise(%)', and 'Revenue', alongside the agent's portfolio state. The action space will include potential trading actions: buy, sell, or hold. Our reward function will be designed to maximize portfolio value over time, encouraging the agent to learn profitable trading strategies. The EDA conducted has demonstrated the potential of these features to inform stock price movements, suggesting that a DRL agent could leverage these relationships to make informed trading decisions. I will split our data into training and testing sets. I will train the model using historical training data and evaluate its performance using a testing data.
These models will enable us to address different aspects of stock market prediction and trading optimization. The supervised learning model will focus on predicting future prices, while the RL model will aim at making profitable trading decisions.
